🍔 Friday - AI Wrap-up #14

AuthorAuthor

Martin Musiol & Daniel Aguspina
July 26, 2024

Llama 3.1 paper is wonderful, with lots of great detail, but the main takeaway […] is that there is nothing particularly special in Anthropic/OpenAI models.

Emad Mostaque, Founder Stability AI

👋 Hey there,

it’s a sunny Friday, and the weekend is around the corner. 🍺 ¯\_(ツ)_/¯  

But first, let’s look at how crazy AI is progressing.

In a nutshell, Open-source models have caught up, people are building amazing products, you will learn it in my upcoming course & we have a landing page now, and breakthrough AI math performance.

Reading time is precisely 145 sec; dai, dai, dai!

own

The gap in performance between closed-source (CS) and open-source (OS) AI has closed due to recent OS AI releases

However, there’s a problem: top performance means big models. Big models mean you need a lot of computing power. Deploying large OS AI (i.e., their weights) is thus impossible without large computing instances.

The aspect of the democratization of AI through OS is thus not valid anymore.

OS AI model releases this week alone

  1. Meta’s Llama 3.1-405B: read about it in our last episode on Wednesday.

Source = https://mail.generativeai.net/p/everyone-can-code

Generative AI - Short & Sweet’s last episode.

  1. Mistral AI’s Large 2: read about it here. It’s not fully open-source. It is released under the Mistral Research License, allowing usage and modification only for research and non-commercial purposes.

    Astonishing: Mistral AI’s team size is in the 11-50 people range. 2-3 orders of magnitude smaller than its competitors 🤯 

The published results are stunning, and the first tests turned out to back this up for me.

Everyone can code!✨ [Course loading]

Speaking of open-source (OS) models, we will build AI products that incorporate OS (and CS) models into our solutions in the course.

AI will do all of this with us in the driver's seat.

Also, we would like to let you know that our course landing page is live!

If you visit + subscribe, or share this page with others, it would mean the world to me. 💝 —> Course Landing Page

Source:https://generativeai.net/everyone-can-code

People are building amazing products with OS AI (i.e., Llama 3.1) in combination with Groq (responsible for AI’s speed)

Example 1: it is impossible to interact with other AI models in that manner.

Example 2: most seamless voice interaction with AI

AI achieves silver-medal standard solving International Mathematical Olympiad problems - unparalleled before

(Source)

AGI, with advanced mathematical reasoning, will unlock new frontiers in science and technology.

Google DeepMind has made great progress and built an AI system.

It combines AlphaProof, a new breakthrough model for formal reasoning, and AlphaGeometry 2, an improved version of our previous system. (I already wrote about it in an earlier episode.)

https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/?utm_source=x&utm_medium=social&utm_campaign=&utm_content=

A formalizer network translates around 1M informal math problems into formal language. A solver network then seeks proofs or disproofs, using the AlphaZero algorithm to progressively tackle more challenging problems.

own

It’s a wrap.

I wish you an awesome weekend.

-Martin